Image Compression and Decompression Using Block Prediction

1. A method performed by a computer for compressing a digital image, the method comprising: identifying, in the image, a target block to be encoded; identifying, in the image, candidate source blocks for predicting the target block, the identifying comprising: clustering blocks of the image by performing hierarchical k-means clustering of the blocks; identifying a cluster obtained by the hierarchical k-means clustering, the cluster including the target block; and selecting one or more blocks from the identified cluster as the candidate source blocks for predicting the target block; selecting a reference block from among the candidate source blocks based on a number of bits required for representing residual information describing a difference between the reference block and the target block; encoding the target block using the reference block; and storing a compressed image having the encoded target block.

2. The method of claim 1 , further comprising: ranking the candidate source blocks in an order; and selecting the reference block based on a number of bits used to specify a position of the reference block in the ranked order.

3. The method of claim 2 , wherein encoding the target block comprises: computing the residual information between the reference block and the target block; and storing the residual information and the position of the reference block in the compressed image.

4. The method of claim 2 , wherein the reference block is selected from among the candidate source blocks using a cost metric based on a number of bits used to represent the residual information and the position, and wherein the cost metric for the selected candidate source block is lower than cost metrics of other candidate source blocks.

5. The method of claim 1 , wherein the cluster including the target block is a leaf-node of a hierarchical k-means tree obtained by the hierarchical k-means clustering.

6. The method of claim 2 , wherein ranking the candidate source blocks in an order comprises: identifying a target region adjacent to the target block; identifying a plurality of candidate source regions, each candidate source region adjacent to a candidate source block; determining differences between the candidate source regions and the target region; and ranking the candidate source blocks in an order determined responsive to the determined differences between the candidate source regions and the target region.

7. The method of claim 6 , wherein determining differences between the candidate source regions and the target region comprises: for each candidate source region, determining a pixel difference between the candidate source region and the target region.

8. The method of claim 1 , wherein selecting the reference block from among the candidate source blocks comprises: for a candidate source block, determining residual information indicating how well the candidate source block predicts the target block; computing a feature vector of the residual information; computing the L1 or L2 norm for the feature vector; and determining whether to select the candidate source block responsive to a cost metric that determines a cost of the candidate source block using an amount of residual information indicating how well the candidate source block predicts the target block and the L1 or L2 norm of the feature vector.

9. The method of claim 8 , wherein cost metrics are determined for the candidate source blocks and wherein the candidate source block is selected responsive to having a minimum cost metric of the cost metrics.

10. A method performed by a computer system for decoding a compressed digital image, the method comprising: identifying, in the image, a target block to be decoded; identifying, in the image, a set of candidate source blocks for predicting the target block, the identifying comprising: clustering blocks of the image by performing hierarchical k-means clustering of the blocks; identifying a cluster obtained by the hierarchical k-means clustering, the cluster including the target block; and selecting one or more blocks from the identified cluster as the candidate source blocks for predicting the target block; retrieving, from the image, a candidate source block from among the identified set of candidate source blocks, wherein the candidate source block is a reference block that was selected during compression of the digital image based on a number of bits used for representing residual information describing a difference between the reference block and the target block; decoding the target block using the reference block; and storing an uncompressed image having the decoded target block.

11. The method of claim 10 , wherein the cluster including the target block is a leaf-node of a hierarchical k-means tree obtained by the hierarchical k-means clustering.

12. The method of claim 10 , further comprising: identifying a target region adjacent to the target block; identifying a plurality of candidate source regions, each source region adjacent to a candidate source block; and determining differences between the candidate source regions and the target region.

13. The method of claim 12 , wherein determining differences between the candidate source regions and the target region comprises: for each candidate source region, determining a pixel difference between the candidate source region and the target region.

14. The method of claim 10 , wherein decoding the target block comprises combining the reference block with the residual information.

15. A computer-implemented system for compressing a digital image, the system comprising: a computer processor; and a non-transitory computer-readable storage medium storing computer program modules configured to execute on the computer processor, the computer program modules comprising: a target block selector module configured to: identify, in an image, a target block to be encoded; a reference block selector module configured to: identify, in the image, candidate source blocks for predicting the target block, the identifying causing the computer processor to: cluster the blocks of the image by performing hierarchical k-means clustering of the blocks; identify a cluster obtained by the hierarchical k-means clustering, the cluster including the target block; and select one or more blocks from the identified cluster as the candidate source blocks for predicting the target block; select a reference block from among the candidate source blocks, the reference block selected based on a number of bits required for representing residual information describing a difference between the reference block and the target block; and a block encoder module configured to: encode the target block using the reference block; and store a compressed image having the encoded target block.

16. The computer-implemented system of claim 15 , wherein the reference block is selected from among the candidate source blocks based on a cost metric indicating a number of bits used to represent the residual information, and wherein the cost metric for the selected candidate source block is lower than cost metrics of other candidate source blocks.

17. The computer-implemented system of claim 15 , wherein the cluster including the target block is a leaf-node of a hierarchical k-means tree obtained by the hierarchical k-means clustering.